Training-Free Spectral Fingerprints of Voice Processing in Transformers

TL;DR

The paper presents a training-free spectral analysis of attention-induced graphs in transformers to detect model-imprinted computational fingerprints during voice alternation. By tracking the Fiedler value $\lambda_2$ and its change $\Delta\lambda_2$ across an early window of layers $2$–$5$, the authors show language- and model-family-specific spectral signatures that correlate with behavior and respond to targeted attention head ablations. Across 20 languages and three model families, they reveal English-specific brittleness in Phi-3-Mini, distributed small effects in Qwen2.5-7B, and systematic moderate effects in LLaMA-3.2-1B, linking tokenizer stress to spectral patterns and demonstrating causal effects via ablations. The framework generalizes beyond voice alternation to reasoning strategies and offers a practical, training-free diagnostic tool for auditing architectural biases and aiding AI safety analyses.

Abstract

Different transformer architectures implement identical linguistic computations via distinct connectivity patterns, yielding model imprinted ``computational fingerprints'' detectable through spectral analysis. Using graph signal processing on attention induced token graphs, we track changes in algebraic connectivity (Fiedler value, $Δλ_2$) under voice alternation across 20 languages and three model families, with a prespecified early window (layers 2--5). Our analysis uncovers clear architectural signatures: Phi-3-Mini shows a dramatic English specific early layer disruption ($\overline{Δλ_2}_{[2,5]}\!\approx\!-0.446$) while effects in 19 other languages are minimal, consistent with public documentation that positions the model primarily for English use. Qwen2.5-7B displays small, distributed shifts that are largest for morphologically rich languages, and LLaMA-3.2-1B exhibits systematic but muted responses. These spectral signatures correlate strongly with behavioral differences (Phi-3: $r=-0.976$) and are modulated by targeted attention head ablations, linking the effect to early attention structure and confirming functional relevance. Taken together, the findings are consistent with the view that training emphasis can leave detectable computational imprints: specialized processing strategies that manifest as measurable connectivity patterns during syntactic transformations. Beyond voice alternation, the framework differentiates reasoning modes, indicating utility as a simple, training free diagnostic for revealing architectural biases and supporting model reliability analysis.

Figures (13)

• Figure 1: Graph Signal Processing framework for transformer analysis. Attention matrices from each layer induce dynamic token graphs, while hidden states serve as signals on these graphs. Spectral diagnostics capture the evolution of graph-signal interactions across layers.
• Figure 2: Phi-3-Mini: strong analytic (EN) negative; otherwise muted.
• Figure 3: Qwen2.5–7B: early–window $\overline{\Delta\lambda_2}_{[2,5]}$ by voice type (mean across languages). Several periphrastic/affixal effects are FDR–sig and negative.
• Figure 4: LLaMA-3.2-1B: muted effects; same ordering trend at lower scale.
• Figure 5: Comparison of spectral diagnostics across syntactic conditions. The Fiedler value $\lambda_2$ shows the most consistent and pronounced differences between layer 2--5.